| Unmanned aerial vehicles(UAVs)are more and more widely used in a variety of scenarios,and object detection technology has brought great improvement to the automation level of UAV tasks,which is the key technology to achieve less manpower to complete complex tasks.With the improvement of hardware equipment performance,the image taken by UAV has the characteristics of small object size,high resolution and high object density,so higher requirements are put forward for the detection of small objects.Small objects contain less information and are easy to disappear in the process of downsampling,so small object detection is still a pain point and difficulty.Inspired by the existence of object recognition assisted by attention mechanism in biological visual mechanism,we propose a new small object detection algorithm.In order to solve the problem that the object is small and easy to be confused with the background,thesis introduces an attention mechanism in the object detection network to retain the information of small objects.In order to solve the problem of vanishing of down-sampled object features,thesis designs a small object detection head,which fuses shalow small object features and deep semantic information in the network to better detect small objects.Based on the above technology,thesis proposes CBAM Micro Head YOLOv5 small object detection network.In addition,thesis also designs a small object detection method based on slice and concat.The research content of thesis mainly includes the following three parts.In the first part,datasets Vis Drone and Tiny Person,which contain a large number of small objects,are selected for dataset preprocessing.The feature analysis of Vis Drone dataset is carried out to observe the detailed statistical information of the data.Finally,the expansion method of Tiny Person small object dataset is briefly introduced.The results show that in the small object dataset,according to a variety of object scale division standards,most of the objects are concentrated in the small scale.In the second part,for the above UAV small object data set,based on the biological attention mechanism and the characteristics of small objects,thesis proposes an improved method,and designs the CBAM Micro Head YOLOv5 object detection network based on the YOLOv5 model.In addition,in order to detect the high-resolution images acquired from the UAV,we propose a small object detection method in a slice and concat method.Based on the above methods,a complete small object detection algorithm is proposed.In the third part,appropriate evaluation indexes are selected to verify the effectiveness of each module proposed in this paper and the slice and concat detection method through experiments.Experiments based on Vis Drone dataset show that the method improves the YOLOv5 s model by 10.9% to 29.9% and improves the YOLOv5 m model by 10.5% to 32.3%.In addition,we conduct experiments in the scene of maritime life search and rescue,and design a data augmentation method for the small number of object samples in some images.Experiments based on Tiny Person dataset show that the data augmentation method improves the YOLOv5 s model by 17.4%,reaching 72.6%,which is a significant improvement. |
